Your First Prompt: Asking on Purpose

A prompt is just the text you hand the model

When you use a large language model, everything you type is the prompt — the input the model reads before it writes anything back. Prompting is the craft of choosing that text well. There is no hidden control panel: the words are the controls. A model that seems brilliant on one phrasing and clueless on another usually saw two different prompts, not two different moods.

Under the hood the model only ever does one thing — predict the next token given everything so far (next-token prediction). Your prompt sets up that continuation. So the real question of prompting is: *what context makes the most useful continuation the most likely one?*

P(x_t \mid x_1,\dots,x_{t-1}) = \operatorname{softmax}\!\left(W\,h_{t-1}\right)

All the model ever computes: a probability for each possible next token, given everything in your prompt so far — your prompt sets up that continuation.

Zero-shot: ask without showing examples

The simplest prompt gives an instruction and no examples at all. This is zero-shot prompting — you rely on what the model already learned during training to do a task it was never explicitly shown by you. Modern instruction-tuned models are surprisingly good at zero-shot for everyday tasks: summarise this, translate that, classify this review.

Classify the sentiment of this review as positive, negative, or neutral.
Respond with only one word.

Review: "The battery lasts forever, but the screen scratches easily."

A zero-shot classification prompt: clear task, clear output format, no examples.

Notice two things that already make this prompt good: it names the allowed labels, and it constrains the output ('only one word'). Zero-shot works best when the task is common and you spell out exactly what a correct answer looks like.

Clarity is the whole game

Instruction clarity does more for your results than any clever trick. The model cannot read your mind; it can only read your words. Vague asks invite the model to guess what you meant, and it will guess plausibly but not necessarily correctly.

State the task as an explicit instruction, not a hint. 'Rewrite this in plain English' beats 'this is too complicated'.
Say who the reader is and how long the answer should be. 'Three sentences for a beginner' removes a dozen guesses.
Specify the output shape — a list, JSON, a single label, a table — so you do not have to reformat by hand.
Add one example of the form you want if words alone feel ambiguous. (We go deeper on examples in the next guide.)

Roles and structure: system vs. user text

Chat models split the prompt into roles. A system message sets durable behaviour — tone, persona, hard rules — and a user message carries the specific request. Putting standing instructions in the system prompt keeps them stable across a conversation instead of repeating them every turn.

Inside a single message, structure still matters. Use delimiters and formatting — headings, triple quotes, XML-like tags — to separate your instruction from the data the model should act on. This stops the model from confusing the text it should follow with the text it should process.

System: You are a careful editor. Fix grammar only; never change meaning.

User:
Instruction: Correct the grammar of the text between the triple quotes.
\"\"\"
Me and him goes to the market every Sundays.
\"\"\"

Roles plus delimiters: the rule lives in the system message; the data is fenced off in the user message.

Why the same idea, worded twice, can differ

Early on you will notice that two prompts you consider equivalent can produce different answers. That is prompt sensitivity: small changes in wording, order, or formatting can shift the output. It is not a bug you can scold away — it falls out of how next-token prediction works. The practical response is to treat your prompt as something you test and refine, not something you write once and trust forever.